Using unsupervised machine learning for automatic entity resolution of natural language records

Abstract

A computer process for entity resolution of natural language records including training a semantic embedding function on a corpus of unlabeled training materials. The semantic embedding function can take a word and represent it as a vector, where the vector represents the word as it relates to the semantic information of the corpus of unlabeled training materials. The process may transform a list of normalized descriptions using the semantic embedding function into a list of vector representations of the descriptions. The process may transform words from a natural language record to a vector representation of the natural language record using the semantic embedding function, and may use a named entity recognizer. The process may find a best match description from the list of normalized descriptions using the list of vector representations of the descriptions and the vector representation of the natural language record, and may include using word mover distance.

What is claimed is: 
     
       1. A computer process for entity resolution of natural language records comprising:
 training an unsupervised semantic embedding function on a corpus of unlabeled training materials using an unsupervised process, where the unsupervised semantic embedding function takes a word and represents it as a semantic vector, where the semantic vector represents the word as it relates to the semantic information of the corpus of unlabeled training materials; 
 generating a list of descriptions using the unsupervised semantic embedding function into a list of description semantic vectors representing the list of descriptions; 
 using words from a natural language record and the unsupervised semantic embedding function to generate a record semantic vector representing the natural language record; and 
 finding a best match description from the list of descriptions by finding a best match description semantic vector from 
 the list of description semantic vectors generated by the unsupervised semantic embedding function trained on the corpus of unlabeled training materials using the unsupervised process that is closest to 
 the record semantic vector generated by the unsupervised semantic embedding function trained on the corpus of unlabeled training materials using the unsupervised process, 
 where finding the best match description uses a screening step that screens the list of descriptions using the description semantic vector from the unsupervised semantic embedding function to get a list of candidate descriptions and a ranking step that further processes the list of candidate descriptions to find the best match description. 
 
     
     
       2. The computer process for entity resolution of natural language records of  claim 1  where the screening step uses a k-d tree search. 
     
     
       3. The computer process for entity resolution of natural language records of  claim 1  where finding the best match uses word mover distance. 
     
     
       4. The computer process for entity resolution of natural language records of  claim 1  further comprising using a named entity recognizer on the natural language record. 
     
     
       5. The computer process for entity resolution of natural language records of  claim 1  further comprising using chunking or collocation detection on the natural language record. 
     
     
       6. The computer process for entity resolution of natural language records of  claim 1  where the training of the semantic embedding function uses a distributional hypothesis. 
     
     
       7. A natural language record entity resolution process comprising:
 using machine learning on a computer to extract semantic information from a corpus of unlabeled training material; 
 generating a list of normalized descriptions using the semantic information into a corresponding set of description semantic math-representations; 
 generating a natural language record using the semantic information to a record semantic math-representation; and 
 determining a best match normalized description for the natural language record using a search process by finding 
 a best match description semantic math-representation from 
 the set of description semantic math-representations that was generated using the extract semantic information from the corpus of unlabeled training material that is closest to 
 the record semantic math-representation that was transformed using the extract semantic information from the corpus of unlabeled training material, 
 where determining the best match normalized description semantic math representation uses a screening step that gets a list of candidate description vectors and a ranking step to further process the list of candidate description vectors to find the best normalized description. 
 
     
     
       8. The natural language record entity resolution process of  claim 7  where the ranking step uses word mover distance. 
     
     
       9. The natural language record entity resolution process of  claim 7  where using machine learning to extract semantic information uses a distributional hypothesis. 
     
     
       10. The computer process for entity resolution of natural language records of  claim 1  where the corpus of unlabeled training materials is selected from a group consisting of journal articles or books. 
     
     
       11. The computer process for entity resolution of natural language records of  claim 1  where the corpus of unlabeled training materials fails to include the list of descriptions or natural language records. 
     
     
       12. The computer process for entity resolution of natural language records of  claim 1  further involves consolidating the record semantic vector of the natural language record to arrive at a consolidated record vector. 
     
     
       13. The computer process for entity resolution of natural language records of  claim 12  where the consolidating the record semantic vector is done by averaging each of the dimensions to arrive at a centroid. 
     
     
       14. The computer process for entity resolution of natural language records of  claim 1  where the list of descriptions includes medical diagnostic descriptions categories.